Predictive Modelling – NMC Conditions

Nowcasting and short-term forecasts for notifiable conditions using EpiNow2

Overview

This section showcases nowcasting and short-term outbreak forecasting for key notifiable medical conditions (NMC) reported through the national surveillance system. The models use the EpiNow2 R package, which combines:

Nowcasting – adjusting for reporting delays
Estimation of $R_t$ – the time-varying effective reproduction number
Short-term forecasting – projecting cases up to 3 weeks ahead

Conditions covered

NMC Category	Conditions
Vaccine-preventable	Fever-Rash (Measles + Rubella), Pertussis, Diphtheria, Tetanus, Hepatitis B, Haemophilus influenzae type B, Congenital rubella syndrome, Poliomyelitis / AFP, Yellow fever
Category 1 — Immediate notification (non-VPD)	Cholera, Malaria, Covid-19, Enteric fever, Meningococcal disease, Listeriosis, Mpox
Category 2 — Routine notification	Hepatitis A

Forecasts are only generated for conditions with ≥ 60 days of data.

Last forecast run

20 February 2026 23:54 — Horizon: 28 days — Conditions: Acute flaccid paralysis, Cholera, Congenital rubella syndrome, Covid-19, Diphtheria, Enteric fever (typhoid or paratyphoid fever), Haemophilus influenzae type B, Hepatitis A, Hepatitis B, Listeriosis, Malaria, Fever-Rash, Meningococcal disease, Mpox, Pertussis, Poliomyelitis, Tetanus, Yellow fever

Surveillance Epicurves

An epidemic curve of reported cases for each forecasted NMC condition, aggregated weekly at the national level.

Figure 1: Weekly reported cases for forecasted NMC conditions

EpiNow2 Forecasts

Each panel shows the median estimate (solid line), 50 % credible interval (dark ribbon) and 90 % credible interval (light ribbon), all smoothed with a 7-day rolling average to remove weekend reporting artefacts. The shaded forecast horizon extends 28 days (4 weeks) beyond the last data point.

Standard EpiNow2 Diagnostic Plots

The panels below show the native EpiNow2 diagnostic output for each condition. Each plot combines three key elements in a single view:

Reported cases — observed data with nowcast / forecast and credible intervals
Estimated infections — latent infection trajectory inferred by the model
Effective reproduction number ($R_t$) — with threshold line at $R_t = 1$

These are produced directly by EpiNow2::plot() and provide a comprehensive snapshot of the epidemic state without any post-processing.

Animated Forecast Outlook

The rolling forecast shows how the predicted trajectory shifted each week as new data arrived. Each coloured line represents a forecast made on a different snapshot date; vertical markers show where each snapshot’s data ended.

This visualisation helps assess:

Forecast stability – are successive predictions consistent?
Outbreak acceleration – are newer forecasts trending higher?
Early warning – divergence between snapshots signals emerging outbreaks.

Figure 2: Rolling weekly EpiNow2 forecasts (animated)

Interpretation Guide

Table 1: How to read the forecast outputs

Element	What it means	Suggested action
Median line	Most likely trajectory based on current Rt	Monitor closely; compare to previous snapshots
50% CI (dark ribbon)	Central range – 1 in 2 chance the true value falls here	Normal operational uncertainty
90% CI (light ribbon)	Wide range – 9 in 10 chance the true value falls here	If real counts approach the upper bound → escalate
Forecast horizon (dashed red)	Where observed data ends and prediction begins	Predictions degrade further from this line
Rolling snapshots diverging upward	Successive forecasts are increasingly pessimistic	Possible outbreak acceleration → trigger response protocols
Rolling snapshots converging	Outbreak is stabilising or declining	Maintain surveillance; consider de-escalation

Effective Reproduction Number ($R_t$)

The time-varying reproduction number $R_t$ indicates how many secondary cases each infected person generates on average.

$R_t > 1$ → epidemic is growing
$R_t = 1$ → epidemic is stable
$R_t < 1$ → epidemic is declining

Figure 3: Estimated Rt for vaccine-preventable conditions

Growth-Rate Summary Table

Table 2: Latest Rt and growth-rate summary per condition

Condition	Latest Rt	Lower 90%	Upper 90%	Assessment	Date
Acute flaccid paralysis	1.01	0.94	1.17	Growing	2026-03-19
Cholera	1.02	0.79	1.28	Growing	2026-03-19
Congenital rubella syndrome	0.84	0.59	1.04	Declining	2026-03-19
Covid-19	0.30	0.02	1.62	Declining	2026-03-19
Diphtheria	0.99	0.92	1.09	Stable / slow decline	2026-03-19
Enteric fever (typhoid or paratyphoid fever)	1.01	0.86	1.12	Growing	2026-03-19
Fever-Rash	0.76	0.42	1.25	Declining	2026-03-19
Haemophilus influenzae type B	1.00	0.94	1.06	Stable / slow decline	2026-03-19
Hepatitis A	0.98	0.79	1.12	Stable / slow decline	2026-03-19
Hepatitis B	0.96	0.57	1.56	Stable / slow decline	2026-03-19
Listeriosis	0.97	0.67	1.23	Stable / slow decline	2026-03-19
Malaria	0.43	0.19	1.00	Declining	2026-03-19
Meningococcal disease	0.99	0.91	1.06	Stable / slow decline	2026-03-19
Mpox	0.85	0.43	1.10	Declining	2026-03-19
Pertussis	0.94	0.76	1.09	Stable / slow decline	2026-03-19
Poliomyelitis	0.69	0.13	1.43	Declining	2026-03-19
Tetanus	0.93	0.61	1.27	Stable / slow decline	2026-03-19
Yellow fever	0.84	0.21	1.49	Declining	2026-03-19

Provincial Breakdown

Figure 4: Weekly cases by province for forecasted NMC conditions

Outbreak Capacity Indicator

A simple traffic-light display combining $R_t$, recent case trends, and forecast trajectory.

Figure 5: Outbreak capacity indicator per condition

Updating Forecasts

Step-by-step instructions to refresh predictions

Forecasts are not computed live during dashboard rendering (they take ~5–15 min per condition). Instead, a batch script produces cached .rds files that the dashboard reads.

Prerequisites:

R ≥ 4.3 with packages: EpiNow2, data.table, dplyr, lubridate, here, purrr
CmdStan installed (EpiNow2 backend). Install via: cmdstanr::install_cmdstan()
Aggregated surveillance data in data/processed/agg_national.rds (produced by R/prepare_dashboard_data.r)

1. Update the surveillance data

# From the project root
cd /Users/briday/Desktop/SAFETP/CLA/NMC_website/NMC_dashboard
Rscript R/prepare_dashboard_data.r

This reads the latest NMC master dataset and produces:

data/processed/agg_national.rds
data/processed/agg_province.rds
data/processed/agg_district.rds

2. Run the forecast batch

Rscript R/run_forecasts.r

This produces for each vaccine-preventable condition:

File	Contents
`forecasts_<condition>.rds`	Single-point EpiNow2 summary (Rt, cases, infections)
`forecasts_obj_<condition>.rds`	Full EpiNow2 result object (for native `plot()` diagnostics)
`forecasts_rolling_<condition>.rds`	Rolling snapshots for animated outlook
`forecasts_meta.rds`	Run timestamp and parameters

3. Re-render the dashboard

quarto render by-forecasts/index.qmd
# Or render the full site:
quarto render

Configuration

Edit the constants at the top of R/run_forecasts.r to adjust:

Parameter	Default	Description
`HORIZON`	28	Days to forecast ahead (4 weeks)
`LOOKBACK`	360	Days of history for each model run
`N_REWINDS`	8	Number of rolling snapshots
`REWIND_STEP`	7	Days between snapshots

Adding a new condition

Ensure the condition appears in agg_national.rds (via prepare_dashboard_data.r)
Add disease parameters in R/epinow2_functions.r → disease_params() function
Add the condition name to forecast_conditions() in the same file
Re-run Rscript R/run_forecasts.r

Model Assumptions & Parameters

The table below shows the disease-specific parameters fed into each EpiNow2 forecast. These are defined in R/epinow2_functions.r → disease_params() and sourced from peer-reviewed epidemiological literature. The Reference column cites the primary source for each parameter set — check these when validating or updating assumptions.

Table 3: EpiNow2 model parameters by condition (with literature references)

Condition	Generation Time	Incubation Period	Rt Prior	Reporting Delay	What This Means	Reference
Fever-Rash	Gamma(shape~34, rate~2.9, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Fever-Rash to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Lessler et al. (2009) Lancet Infect Dis 9(5):291-300
Measles	Gamma(shape~34, rate~2.9, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Measles to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Lessler et al. (2009) Lancet Infect Dis 9(5):291-300
Rubella	Gamma(shape~27, rate~1.5, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Rubella to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Vink et al. (2014) Am J Epidemiol 180(9):865-875; Lessler et al. (2009)
Pertussis	Gamma(shape~13, rate~0.6, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Pertussis to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Vink et al. (2014); Wearing & Rohani (2009) Emerg Infect Dis 15(8):1248
Diphtheria	Gamma(shape~7, rate~0.9, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Diphtheria to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Truelove et al. (2020) Emerg Infect Dis 26(10):2396; AAP Red Book
Tetanus	Gamma(shape~8, rate~0.6, max=--)	Gamma (see source)	Normal(mean=1.0, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Tetanus to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 1 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	WHO Tetanus Position Paper (2017); Roper et al. (2018) Lancet 391:1657
Hepatitis B	Gamma(shape~9, rate~0.1, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Hepatitis B to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Edmunds et al. (1993) Epidemiol Infect 110(3):569; Liang (2009) Hepatology 49(S5):S13
Haemophilus influenzae type B	Gamma(shape~4, rate~1.3, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Haemophilus influenzae type B to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Peltola (2000) Clin Microbiol Rev 13(2):302-317; AAP Red Book (2024)
Congenital rubella syndrome	Gamma(shape~27, rate~1.5, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Congenital rubella syndrome to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Vink et al. (2014); Lessler et al. (2009) – rubella parameters
Poliomyelitis	Gamma(shape~9, rate~0.6, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Poliomyelitis to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Fine & Carneiro (1999) Am J Epidemiol 150(12):1332; Nathanson & Kew (2010) Am J Epidemiol 172(11):1213
Acute flaccid paralysis	Gamma(shape~9, rate~0.6, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Acute flaccid paralysis to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Fine & Carneiro (1999) Am J Epidemiol 150(12):1332; Nathanson & Kew (2010) Am J Epidemiol 172(11):1213
Yellow fever	Gamma(shape~16, rate~1.3, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Yellow fever to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Johansson et al. (2010) Am J Trop Med Hyg 82(5):930; Monath & Vasconcelos (2015) J Clin Virol 64:160
Cholera	Gamma(shape~6, rate~1.2, max=--)	Gamma (see source)	Normal(mean=2.0, sd=1.0)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Cholera to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Azman et al. (2013) J Infect Dis 207(11):1698-1706
Malaria	Gamma(shape~14, rate~0.5, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Malaria to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Huber et al. (2016) Malar J 15:18; WHO World Malaria Report (2023)
Covid-19	Gamma(shape~6, rate~2.0, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Covid-19 to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Park et al. (2023) BMC Infect Dis 23:295; Wu et al. (2022) JAMA Netw Open 5(8):e2228008
Enteric fever (typhoid or paratyphoid fever)	Gamma(shape~12, rate~0.9, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Enteric fever (typhoid or paratyphoid fever) to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Pitzer et al. (2014) PLoS Negl Trop Dis 8(6):e2891; Crump et al. (2015) Clin Microbiol Rev 28(4):901
Meningococcal disease	Gamma(shape~8, rate~1.1, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Meningococcal disease to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Trotter et al. (2005) Emerg Infect Dis 11(4):563; Rosenstein et al. (2001) NEJM 344(18):1378
Listeriosis	Gamma(shape~4, rate~0.2, max=--)	Gamma (see source)	Normal(mean=1.2, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Listeriosis to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 1 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Goulet et al. (2013) Clin Infect Dis 56(10):1411-1419
Mpox	Gamma(shape~8, rate~0.7, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Mpox to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Miura et al. (2022) Euro Surveill 27(44):2200806; Thornhill et al. (2022) NEJM 387(8):679
Hepatitis A	Gamma(shape~16, rate~0.7, max=--)	Gamma (see source)	Normal(mean=1.5, sd=0.5)	LogNormal(mean=3, sd=2, max=15)	On average, an infected person can spread Hepatitis A to others about several days after being infected themselves. Symptoms typically appear ~several days after exposure. The model starts by assuming each case infects about 2 other people on average, then adjusts this up or down based on the actual reported data. A ~3-day delay is assumed between someone falling ill and the case being officially reported.	Jacobsen & Wiersma (2010) Clin Liver Dis 14(4):591; Lemon et al. (2018) Lancet 391:1587
Generation time ≈ serial interval where direct measurement is unavailable. Incubation period = time from exposure to symptom onset. All distributions are Gamma. Reporting delay: default LogNormal(mean=3, sd=2, max=15) when empirical delay data is unavailable.

Reading the parameters

Parameter	What it controls	Why it matters
Generation time	Average time between successive infections in a chain (≈ serial interval)	Determines how fast $R_t$ changes translate into case-count changes
Incubation period	Time from infection/exposure to symptom onset	Affects the delay between infection dynamics and observed cases
$R_t$ prior	Starting assumption for reproduction number	Weakly informative — the model quickly learns from data
Reporting delay	Time from symptom onset to NMC notification	Accounts for the surveillance lag between illness and official capture
What This Means	Plain-language summary of all parameters for this condition	Helps non-technical readers understand what the model assumes
Reference	Primary literature source	Used for validation — update if newer estimates become available

How to adjust parameters

To update parameter values for a condition:

Open R/epinow2_functions.r
Find the condition in disease_params() (search for the condition name in lowercase)
Change mean, sd, max values to match your preferred literature source
Update the reference string to cite the new source
Re-run: Rscript R/run_forecasts.r
Re-render: quarto render by-forecasts/index.qmd

Example — changing measles generation time to use a new study:

"fever-rash" = , "measles" = list(
  generation_time   = EpiNow2::Gamma(mean = 11.0, sd = 2.5, max = 20),
  # ↑ changed from 11.7 to 11.0 based on New Study (2026)
  incubation_period = EpiNow2::Gamma(mean = 12.5, sd = 2.3, max = 21),
  rt_prior          = list(mean = 2, sd = 1),
  reference         = "New Study (2026) Journal Name DOI:..."
),

Methods & Technical Details

Technical details (click to expand)

Model: EpiNow2 v1.8.0 – a Bayesian semi-mechanistic model that jointly estimates the effective reproduction number ($R_t$), infection incidence, and reporting patterns. The model is fitted using Hamiltonian Monte Carlo via CmdStan. A day-of-week effect is enabled (obs_opts(week_effect = TRUE)) to account for lower weekend reporting.

Display smoothing: All custom forecast plots apply a 7-day rolling average to both the observed surveillance series and the EpiNow2 median / credible-interval outputs. This removes residual day-of-week jaggedness and highlights the underlying epidemic trend.

Reporting delay: When empirical symptom→notification dates are available in the NMC data, the delay distribution is estimated via EpiNow2::estimate_delay() (capped at 15 days, max 100 observations). Otherwise a default LogNormal(mean=3, sd=2, max=15) prior is used.

Generation time & incubation period: Condition-specific Gamma distributions from published peer-reviewed literature. Key sources: Lessler et al. (2009) for measles/rubella; Vink et al. (2014) for serial intervals; Azman et al. (2013) for cholera; Park et al. (2023) for COVID-19 Omicron estimates. See the parameter table above for full citations. These are fixed inputs — not estimated from the data.

Priors: $R_t$ priors are weakly informative: Normal(2, 1) for directly-transmitted pathogens, Normal(1.5, 0.5) for vector-borne / lower-Rt conditions. The model quickly learns from data.

Forecast horizon: 28 days (4 weeks) ahead from the last observation date.

History window: 360 days of surveillance data fed to each model run.

Completeness / Timeliness: The NMC is a passive surveillance system; under-reporting varies by condition and province. Forecasts represent reported cases, not true incidence.

Rolling forecast animation: Re-runs the model 8 times (weekly snapshots), each using data up to that snapshot date, to visualise how the forecast outlook evolved over time.

--- title: "Predictive Modelling – NMC Conditions" subtitle: "Nowcasting and short-term forecasts for notifiable conditions using EpiNow2" format: html: toc: true toc-depth: 3 code-fold: true code-tools: true smooth-scroll: true fig-width: 10 fig-height: 6 execute: echo: false warning: false message: false cache: false --- ```{r} #| label: setup #| include: false # ── Package loading ────────────────────────────────────────────────────────── if (!requireNamespace("pacman", quietly = TRUE)) install.packages("pacman") pacman::p_load( data.table, dplyr, tidyr, lubridate, purrr, stringr, ggplot2, plotly, gganimate, gifski, DT, flextable, scales, here, EpiNow2, patchwork ) # ── Source shared forecasting functions ────────────────────────────────────── source(here::here("R", "epinow2_functions.r")) # ── Paths ──────────────────────────────────────────────────────────────────── data_dir <- here::here("data", "processed") # ── Load aggregated surveillance data ──────────────────────────────────────── agg_national <- readRDS(file.path(data_dir, "agg_national.rds")) setDT(agg_national) # Combine Measles + Rubella → Fever-Rash and filter to forecast-eligible conditions agg_vp <- filter_vp_conditions(agg_national, include_fever_rash = TRUE) agg_fc <- filter_forecast_conditions(agg_national, include_fever_rash = TRUE) # Available conditions in data avail_conditions <- unique(agg_fc$condition) # ── Try to load pre-computed forecasts (from R/run_forecasts.r) ────────────── # If forecast files exist we use them; otherwise we run EpiNow2 live (slower). forecast_files <- list.files(data_dir, pattern = "^forecasts_.*\\.rds$", full.names = TRUE) has_cached <- length(forecast_files) > 0 load_forecast_rds <- function(pattern, exclude_patterns = NULL) { ff <- list.files(data_dir, pattern = pattern, full.names = TRUE) if (!is.null(exclude_patterns)) { for (ep in exclude_patterns) { ff <- ff[!grepl(ep, basename(ff))] } } if (length(ff) == 0) return(NULL) bind_rows(lapply(ff, readRDS)) } forecast_single <- load_forecast_rds( "^forecasts_.*\\.rds$", exclude_patterns = c("^forecasts_rolling_", "^forecasts_meta", "^forecasts_obj_") ) forecast_rolling <- load_forecast_rds("^forecasts_rolling_.*\\.rds$") # Load metadata if available meta_path <- file.path(data_dir, "forecasts_meta.rds") meta <- if (file.exists(meta_path)) readRDS(meta_path) else NULL # ── Colour palette ─────────────────────────────────────────────────────────── condition_colours <- c( # Vaccine-preventable "Fever-Rash" = "#7b2cbf", "Measles" = "#9d4edd", "Rubella" = "#c77dff", "Pertussis" = "#e07a5f", "Diphtheria" = "#3d85c6", "Tetanus" = "#8338ec", "Hepatitis B" = "#ff006e", "Haemophilus influenzae type B" = "#fb5607", "Congenital rubella syndrome" = "#e0aaff", "Poliomyelitis" = "#3a86ff", "Acute flaccid paralysis" = "#457b9d", "Yellow fever" = "#ffbe0b", # Category 1 – high priority (non-VPD) "Cholera" = "#0077b6", "Malaria" = "#2d6a4f", "Covid-19" = "#d90429", "Enteric fever (typhoid or paratyphoid fever)" = "#f4a261", "Meningococcal disease" = "#264653", "Listeriosis" = "#e76f51", "Mpox" = "#606c38", # Category 2 "Hepatitis A" = "#bc6c25" ) # ── 7-day rolling average helper (smooths weekend reporting artefact) ──────── roll7 <- function(x) { n <- length(x) out <- rep(NA_real_, n) for (i in 7:n) out[i] <- mean(x[(i - 6):i], na.rm = TRUE) out } # ── Theme ──────────────────────────────────────────────────────────────────── theme_nmc <- theme_minimal(base_size = 13) + theme( plot.title = element_text(face = "bold", size = 15), plot.subtitle = element_text(colour = "grey40"), legend.position = "bottom", panel.grid.minor = element_blank() ) ``` ## Overview This section showcases **nowcasting and short-term outbreak forecasting** for key notifiable medical conditions (NMC) reported through the national surveillance system. The models use the **EpiNow2** R package, which combines: - **Nowcasting** – adjusting for reporting delays - **Estimation of $R_t$** – the time-varying effective reproduction number - **Short-term forecasting** – projecting cases up to 3 weeks ahead ::: {.callout-note} ### Conditions covered | NMC Category | Conditions | |---|---| | **Vaccine-preventable** | Fever-Rash (Measles + Rubella), Pertussis, Diphtheria, Tetanus, Hepatitis B, Haemophilus influenzae type B, Congenital rubella syndrome, Poliomyelitis / AFP, Yellow fever | | **Category 1** — Immediate notification (non-VPD) | Cholera, Malaria, Covid-19, Enteric fever, Meningococcal disease, Listeriosis, Mpox | | **Category 2** — Routine notification | Hepatitis A | *Forecasts are only generated for conditions with ≥ 60 days of data.* ::: ```{r} #| label: meta-info #| results: asis if (!is.null(meta)) { cat(sprintf( "::: {.callout-tip}\n### Last forecast run\n**%s** — Horizon: %d days — Conditions: %s\n:::\n", format(meta$run_time, "%d %B %Y %H:%M"), meta$horizon, paste(meta$conditions, collapse = ", ") )) } else { cat("::: {.callout-warning}\nNo cached forecasts found. See the **Updating Forecasts** section below to generate them.\n:::\n") } ``` --- ## Surveillance Epicurves {#epicurves} An epidemic curve of reported cases for each forecasted NMC condition, aggregated weekly at the national level. ```{r} #| label: fig-epicurves #| fig-cap: "Weekly reported cases for forecasted NMC conditions" #| fig-height: 10 # Aggregate to weekly agg_weekly <- agg_fc[, .( cases = sum(confirm, na.rm = TRUE) ), by = .(condition, week_start = floor_date(date, "week"))] # Only conditions with > 0 total cases agg_weekly <- agg_weekly[, total := sum(cases), by = condition][total > 0][, total := NULL] p_epi <- ggplot(agg_weekly, aes(x = week_start, y = cases, fill = condition)) + geom_col(alpha = 0.85, width = 6) + facet_wrap(~ condition, scales = "free_y", ncol = 1) + scale_fill_manual(values = condition_colours, guide = "none") + scale_x_date(date_breaks = "3 months", date_labels = "%b %Y") + scale_y_continuous(expand = expansion(mult = c(0, 0.05))) + labs(x = NULL, y = "Weekly cases") + theme_nmc ggplotly(p_epi, tooltip = c("x", "y")) %>% layout(legend = list(orientation = "h", y = -0.1)) ``` --- ## EpiNow2 Forecasts {#forecasts} Each panel shows the **median estimate** (solid line), **50 % credible interval** (dark ribbon) and **90 % credible interval** (light ribbon), all smoothed with a **7-day rolling average** to remove weekend reporting artefacts. The shaded forecast horizon extends **`r if (!is.null(meta)) meta$horizon else 28` days** (4 weeks) beyond the last data point. ::: {.panel-tabset} ```{r} #| label: forecast-tabs #| results: asis # If we have cached single-point forecasts, plot them if (!is.null(forecast_single) && nrow(forecast_single) > 0) { conditions_to_plot <- unique(forecast_single$condition) for (cond in conditions_to_plot) { cat(sprintf("\n### %s\n\n", cond)) df_cond <- forecast_single %>% filter(condition == cond) # Grab matching surveillance time series for backdrop (7-day smoothed) ts_raw <- prepare_condition_ts(agg_fc, cond) ts_raw <- ts_raw %>% filter(date >= max(date) - days(180)) %>% arrange(date) %>% mutate(confirm_smooth = roll7(confirm)) # Determine forecast start (where type changes) forecast_start <- df_cond %>% filter(variable == "reported_cases", grepl("forecast", type)) %>% pull(date) %>% min(na.rm = TRUE) # Smooth the EpiNow2 median & CI with 7-day rolling average df_inf <- df_cond %>% filter(variable == "reported_cases") %>% arrange(date) %>% mutate( median = roll7(median), lower_50 = roll7(lower_50), upper_50 = roll7(upper_50), lower_90 = roll7(lower_90), upper_90 = roll7(upper_90) ) %>% filter(!is.na(median)) p <- ggplot() + # Surveillance: light bars (raw) + smoothed line overlay geom_col(data = ts_raw, aes(x = date, y = confirm), fill = condition_colours[cond] %||% "grey60", alpha = 0.15, width = 1) + geom_line(data = ts_raw %>% filter(!is.na(confirm_smooth)), aes(x = date, y = confirm_smooth), colour = condition_colours[cond] %||% "grey60", alpha = 0.5, linewidth = 0.4) + # 90% CI ribbon geom_ribbon(data = df_inf, aes(x = date, ymin = lower_90, ymax = upper_90), fill = condition_colours[cond] %||% "grey60", alpha = 0.15) + # 50% CI ribbon geom_ribbon(data = df_inf, aes(x = date, ymin = lower_50, ymax = upper_50), fill = condition_colours[cond] %||% "grey60", alpha = 0.3) + # Median line (smoothed) geom_line(data = df_inf, aes(x = date, y = median), colour = condition_colours[cond] %||% "black", linewidth = 0.8) + # Forecast boundary geom_vline(xintercept = forecast_start, linetype = "dashed", colour = "red", alpha = 0.6) + annotate("text", x = forecast_start + days(1), y = Inf, label = "Forecast →", vjust = 2, hjust = 0, colour = "red", size = 3.5) + scale_y_continuous(expand = expansion(mult = c(0, 0.05))) + scale_x_date(date_breaks = "1 month", date_labels = "%b %Y") + labs( title = paste(cond, "— EpiNow2 Forecast (7-day smoothed)"), subtitle = "Median with 50% and 90% credible intervals · 4-week horizon", x = NULL, y = "Daily reported cases (7-day avg)" ) + theme_nmc # Print plotly print(htmltools::tagList(ggplotly(p, tooltip = c("x", "y")))) cat("\n\n") } } else { cat("\n### No forecasts available\n\n") cat("Run `Rscript R/run_forecasts.r` from the project root to generate cached forecast data.\n\n") } ``` ::: --- ## Standard EpiNow2 Diagnostic Plots {#epinow2-diagnostics} The panels below show the **native EpiNow2 diagnostic output** for each condition. Each plot combines three key elements in a single view: 1. **Reported cases** — observed data with nowcast / forecast and credible intervals 2. **Estimated infections** — latent infection trajectory inferred by the model 3. **Effective reproduction number ($R_t$)** — with threshold line at $R_t = 1$ These are produced directly by `EpiNow2::plot()` and provide a comprehensive snapshot of the epidemic state without any post-processing. ::: {.panel-tabset} ```{r} #| label: epinow2-standard-plots #| results: asis #| fig-height: 9 #| fig-width: 10 # Load full EpiNow2 result objects saved by R/run_forecasts.r obj_files <- list.files(data_dir, pattern = "^forecasts_obj_.*\\.rds$", full.names = TRUE) if (length(obj_files) > 0) { for (of in sort(obj_files)) { # Extract condition name from filename cond_slug <- gsub("^forecasts_obj_|\\.rds$", "", basename(of)) cond_label <- gsub("_", " ", cond_slug) |> tools::toTitleCase() cat(sprintf("\n### %s\n\n", cond_label)) tryCatch({ obj <- readRDS(of) p <- plot(obj) print(p + patchwork::plot_annotation( title = paste0(cond_label, " \u2014 EpiNow2 Diagnostic Summary"), subtitle = "Nowcast / forecast \u00b7 estimated infections \u00b7 Rt", theme = theme( plot.title = element_text(face = "bold", size = 15), plot.subtitle = element_text(colour = "grey40", size = 11) ) )) cat("\n\n") }, error = function(e) { cat(sprintf("*Plot could not be rendered for %s: %s*\n\n", cond_label, e$message)) }) } } else { cat("\n### No EpiNow2 result objects found\n\n") cat("Run `Rscript R/run_forecasts.r` to generate full result objects for standard diagnostic plots.\n\n") } ``` ::: --- ## Animated Forecast Outlook {#animation} The rolling forecast shows how the **predicted trajectory shifted** each week as new data arrived. Each coloured line represents a forecast made on a different snapshot date; vertical markers show where each snapshot's data ended. This visualisation helps assess: - **Forecast stability** – are successive predictions consistent? - **Outbreak acceleration** – are newer forecasts trending higher? - **Early warning** – divergence between snapshots signals emerging outbreaks. ```{r} #| label: fig-rolling-forecast #| fig-cap: "Rolling weekly EpiNow2 forecasts (animated)" #| fig-height: 7 if (!is.null(forecast_rolling) && nrow(forecast_rolling) > 0) { # Choose the condition with highest recent case count for the showcase showcase_cond <- forecast_rolling %>% group_by(condition) %>% summarise(total_med = sum(median, na.rm = TRUE)) %>% arrange(desc(total_med)) %>% pull(condition) %>% first() roll_df <- forecast_rolling %>% filter(condition == showcase_cond, variable == "reported_cases") %>% mutate(median = ifelse(median > quantile(median, 0.99, na.rm = TRUE), NA, median)) ts_raw <- prepare_condition_ts(agg_fc, showcase_cond) ts_raw <- ts_raw %>% filter(date >= min(roll_df$date, na.rm = TRUE)) # Static version with all snapshots p_roll <- ggplot() + geom_col(data = ts_raw, aes(x = date, y = confirm), fill = "grey70", alpha = 0.4, width = 1) + geom_line(data = roll_df, aes(x = date, y = median, group = snapshot_label, colour = snapshot_label), linewidth = 0.5, alpha = 0.8) + scale_colour_viridis_d(option = "plasma", name = "Snapshot date") + scale_y_continuous(expand = expansion(mult = c(0, 0.05))) + scale_x_date(date_breaks = "1 month", date_labels = "%b %Y") + labs( title = paste(showcase_cond, "— Rolling Weekly Forecasts"), subtitle = "Each line is a forecast made on a different snapshot date", x = NULL, y = "Daily reported cases" ) + theme_nmc ggplotly(p_roll, tooltip = c("x", "y", "colour")) %>% layout(legend = list(orientation = "h", y = -0.15)) } else { cat("No rolling forecast data available. Run `Rscript R/run_forecasts.r` to generate.\n") } ``` ```{r} #| label: fig-animated-gif #| fig-cap: "Animated forecast evolution — each frame adds a new weekly snapshot" #| dev: "png" #| fig-width: 9 #| fig-height: 5 #| eval: false # Build a gganimate version (rendered as GIF in the HTML page) has_rolling <- !is.null(forecast_rolling) && NROW(forecast_rolling) > 0 if (has_rolling) { # Use the same showcase condition — aggregate to WEEKLY to keep memory low anim_df <- forecast_rolling %>% filter(condition == showcase_cond, variable == "reported_cases") %>% mutate( week = floor_date(date, "week"), snapshot_id = as.integer(factor(snapshot_date)) ) %>% group_by(snapshot_date, snapshot_label, snapshot_id, week) %>% summarise(median = mean(median, na.rm = TRUE), .groups = "drop") %>% filter(!is.na(median)) %>% mutate(median = ifelse(median > quantile(median, 0.99, na.rm = TRUE), NA, median)) ts_bg <- prepare_condition_ts(agg_fc, showcase_cond) %>% filter(date >= min(anim_df$week, na.rm = TRUE)) %>% mutate(week = floor_date(date, "week")) %>% group_by(week) %>% summarise(confirm = sum(confirm, na.rm = TRUE), .groups = "drop") n_snapshots <- length(unique(anim_df$snapshot_id)) p_anim <- ggplot() + geom_col(data = ts_bg, aes(x = week, y = confirm), fill = "grey75", alpha = 0.4, width = 6) + geom_line(data = anim_df, aes(x = week, y = median, group = snapshot_label, colour = snapshot_label), linewidth = 0.7) + scale_colour_viridis_d(option = "plasma", name = "Snapshot") + scale_y_continuous(expand = expansion(mult = c(0, 0.05))) + scale_x_date(date_breaks = "1 month", date_labels = "%b %Y") + labs( title = paste(showcase_cond, "forecast as of: {closest_state}"), subtitle = "Cumulative forecast snapshots – watch the outlook evolve", x = NULL, y = "Weekly reported cases" ) + theme_nmc + transition_states(snapshot_label, transition_length = 1, state_length = 2) + enter_fade() + exit_fade() tryCatch({ animate(p_anim, nframes = n_snapshots * 4, fps = 4, width = 700, height = 400, res = 96, renderer = gifski_renderer()) }, error = function(e) { cat("Animation could not be rendered:", conditionMessage(e), "\n") }) } else { cat("No rolling forecast data available — animation skipped.\n") } ``` --- ## Interpretation Guide {#interpretation} ```{r} #| label: tbl-interpretation #| tbl-cap: "How to read the forecast outputs" tribble( ~Element, ~Meaning, ~`Action if concerning`, "Median line", "Most likely trajectory based on current Rt", "Monitor closely; compare to previous snapshots", "50% CI (dark ribbon)", "Central range – 1 in 2 chance the true value falls here", "Normal operational uncertainty", "90% CI (light ribbon)", "Wide range – 9 in 10 chance the true value falls here", "If real counts approach the upper bound → escalate", "Forecast horizon (dashed red)", "Where observed data ends and prediction begins", "Predictions degrade further from this line", "Rolling snapshots diverging upward", "Successive forecasts are increasingly pessimistic", "Possible outbreak acceleration → trigger response protocols", "Rolling snapshots converging", "Outbreak is stabilising or declining", "Maintain surveillance; consider de-escalation" ) %>% flextable() %>% set_header_labels( Element = "Element", Meaning = "What it means", `Action if concerning` = "Suggested action" ) %>% theme_vanilla() %>% autofit() %>% fontsize(size = 10, part = "all") %>% bold(part = "header") ``` --- ## Effective Reproduction Number ($R_t$) {#rt} The time-varying reproduction number $R_t$ indicates **how many secondary cases** each infected person generates on average. - $R_t > 1$ → epidemic is **growing** - $R_t = 1$ → epidemic is **stable** - $R_t < 1$ → epidemic is **declining** ```{r} #| label: fig-rt #| fig-cap: "Estimated Rt for vaccine-preventable conditions" #| fig-height: 6 if (!is.null(forecast_single) && nrow(forecast_single) > 0) { rt_df <- forecast_single %>% filter(variable == "R") %>% filter(!is.na(median)) if (nrow(rt_df) > 0) { p_rt <- ggplot(rt_df, aes(x = date)) + geom_ribbon(aes(ymin = lower_90, ymax = upper_90, fill = condition), alpha = 0.15) + geom_ribbon(aes(ymin = lower_50, ymax = upper_50, fill = condition), alpha = 0.3) + geom_line(aes(y = median, colour = condition), linewidth = 0.8) + geom_hline(yintercept = 1, linetype = "dashed", colour = "red") + annotate("text", x = min(rt_df$date), y = 1.05, label = "Rt = 1 (epidemic threshold)", hjust = 0, colour = "red", size = 3) + facet_wrap(~ condition, scales = "free_y", ncol = 1) + scale_fill_manual(values = condition_colours, guide = "none") + scale_colour_manual(values = condition_colours, guide = "none") + scale_x_date(date_breaks = "1 month", date_labels = "%b %Y") + labs( title = "Effective Reproduction Number (Rt)", subtitle = "Median with 50% and 90% credible intervals", x = NULL, y = expression(R[t]) ) + theme_nmc ggplotly(p_rt, tooltip = c("x", "y")) %>% layout(legend = list(orientation = "h", y = -0.1)) } else { cat("No Rt estimates available in the cached forecasts.\n") } } else { cat("Run `Rscript R/run_forecasts.r` to generate Rt estimates.\n") } ``` --- ## Growth-Rate Summary Table {#growth-table} ```{r} #| label: tbl-growth #| tbl-cap: "Latest Rt and growth-rate summary per condition" if (!is.null(forecast_single) && nrow(forecast_single) > 0) { # Latest Rt per condition latest_rt <- forecast_single %>% filter(variable == "R", !is.na(median)) %>% group_by(condition) %>% filter(date == max(date)) %>% ungroup() %>% transmute( Condition = condition, `Latest Rt` = round(median, 2), `Lower 90%` = round(lower_90, 2), `Upper 90%` = round(upper_90, 2), `Assessment` = case_when( median > 1.5 ~ "Rapid growth ⚠️", median > 1 ~ "Growing", median > 0.9 ~ "Stable / slow decline", TRUE ~ "Declining" ), `Date` = date ) flextable(latest_rt) %>% theme_vanilla() %>% autofit() %>% fontsize(size = 10, part = "all") %>% bold(part = "header") %>% color(i = ~ `Latest Rt` > 1.5, j = "Assessment", color = "#b91c1c") %>% color(i = ~ `Latest Rt` > 1 & `Latest Rt` <= 1.5, j = "Assessment", color = "#b45309") %>% color(i = ~ `Latest Rt` <= 1, j = "Assessment", color = "#065f46") } else { cat("Forecast data not yet generated.\n") } ``` --- ## Provincial Breakdown {#provincial} ```{r} #| label: fig-provincial #| fig-cap: "Weekly cases by province for forecasted NMC conditions" #| fig-height: 10 # Use raw aggregated data (province level) agg_prov_path <- file.path(data_dir, "agg_province.rds") if (file.exists(agg_prov_path)) { agg_prov <- readRDS(agg_prov_path) setDT(agg_prov) agg_prov_fc <- filter_forecast_conditions(agg_prov, include_fever_rash = TRUE) agg_prov_fc[, week_start := floor_date(date, "week")] prov_weekly <- agg_prov_fc[, .(cases = sum(confirm, na.rm = TRUE)), by = .(condition, prov_, week_start)] prov_weekly <- prov_weekly[cases > 0] if (nrow(prov_weekly) > 0) { p_prov <- ggplot(prov_weekly, aes(x = week_start, y = cases, fill = prov_)) + geom_col(alpha = 0.8, width = 6) + facet_wrap(~ condition, scales = "free_y", ncol = 2) + scale_x_date(date_breaks = "3 months", date_labels = "%b %Y") + scale_y_continuous(expand = expansion(mult = c(0, 0.05))) + labs( title = "Provincial distribution of forecasted NMC conditions", x = NULL, y = "Weekly cases", fill = "Province" ) + theme_nmc ggplotly(p_prov, tooltip = c("x", "y", "fill")) %>% layout(legend = list(orientation = "h", y = -0.1)) } else { cat("No provincial data available for the selected conditions.\n") } } else { cat("Provincial aggregated data not found. Run `Rscript R/prepare_dashboard_data.r` first.\n") } ``` --- ## Outbreak Capacity Indicator {#capacity} A simple traffic-light display combining $R_t$, recent case trends, and forecast trajectory. ```{r} #| label: fig-capacity #| fig-cap: "Outbreak capacity indicator per condition" if (!is.null(forecast_single) && nrow(forecast_single) > 0) { # For each condition: latest Rt + 7-day case trend capacity <- forecast_single %>% filter(variable == "R", !is.na(median)) %>% group_by(condition) %>% filter(date == max(date)) %>% ungroup() %>% select(condition, rt_median = median) %>% mutate( signal = case_when( rt_median > 1.5 ~ "High", rt_median > 1.0 ~ "Moderate", TRUE ~ "Low" ), colour = case_when( signal == "High" ~ "#dc2626", signal == "Moderate" ~ "#f59e0b", signal == "Low" ~ "#16a34a" ) ) p_cap <- ggplot(capacity, aes(x = reorder(condition, rt_median), y = rt_median, fill = colour)) + geom_col(width = 0.6) + geom_hline(yintercept = 1, linetype = "dashed", colour = "grey40") + geom_text(aes(label = paste0("Rt = ", round(rt_median, 2), "\n", signal)), vjust = -0.3, size = 3.5) + scale_fill_identity() + coord_flip() + labs( title = "Outbreak Capacity Indicator", subtitle = "Based on latest estimated Rt", x = NULL, y = expression(R[t]) ) + theme_nmc + theme(panel.grid.major.y = element_blank()) ggplotly(p_cap, tooltip = c("y", "text")) } else { cat("Forecast data required. See **Updating Forecasts** below.\n") } ``` --- ## Updating Forecasts {#updating} ::: {.callout-important} ### Step-by-step instructions to refresh predictions Forecasts are **not computed live** during dashboard rendering (they take ~5–15 min per condition). Instead, a batch script produces cached `.rds` files that the dashboard reads. **Prerequisites:** 1. R ≥ 4.3 with packages: `EpiNow2`, `data.table`, `dplyr`, `lubridate`, `here`, `purrr` 2. CmdStan installed (EpiNow2 backend). Install via: `cmdstanr::install_cmdstan()` 3. Aggregated surveillance data in `data/processed/agg_national.rds` (produced by `R/prepare_dashboard_data.r`) ::: ### 1. Update the surveillance data ```bash # From the project root cd /Users/briday/Desktop/SAFETP/CLA/NMC_website/NMC_dashboard Rscript R/prepare_dashboard_data.r ``` This reads the latest NMC master dataset and produces: - `data/processed/agg_national.rds` - `data/processed/agg_province.rds` - `data/processed/agg_district.rds` ### 2. Run the forecast batch ```bash Rscript R/run_forecasts.r ``` This produces for each vaccine-preventable condition: | File | Contents | |---|---| | `forecasts_<condition>.rds` | Single-point EpiNow2 summary (Rt, cases, infections) | | `forecasts_obj_<condition>.rds` | Full EpiNow2 result object (for native `plot()` diagnostics) | | `forecasts_rolling_<condition>.rds` | Rolling snapshots for animated outlook | | `forecasts_meta.rds` | Run timestamp and parameters | ### 3. Re-render the dashboard ```bash quarto render by-forecasts/index.qmd # Or render the full site: quarto render ``` ### Configuration Edit the constants at the top of `R/run_forecasts.r` to adjust: | Parameter | Default | Description | |---|---|---| | `HORIZON` | 28 | Days to forecast ahead (4 weeks) | | `LOOKBACK` | 360 | Days of history for each model run | | `N_REWINDS` | 8 | Number of rolling snapshots | | `REWIND_STEP` | 7 | Days between snapshots | ### Adding a new condition 1. Ensure the condition appears in `agg_national.rds` (via `prepare_dashboard_data.r`) 2. Add disease parameters in `R/epinow2_functions.r` → `disease_params()` function 3. Add the condition name to `forecast_conditions()` in the same file 4. Re-run `Rscript R/run_forecasts.r` --- ## Model Assumptions & Parameters {#assumptions} The table below shows the **disease-specific parameters** fed into each EpiNow2 forecast. These are defined in `R/epinow2_functions.r` → `disease_params()` and sourced from peer-reviewed epidemiological literature. The **Reference** column cites the primary source for each parameter set — check these when validating or updating assumptions. ```{r} #| label: tbl-assumptions #| tbl-cap: "EpiNow2 model parameters by condition (with literature references)" #| cache: false # Build the assumptions table from the helper function assumptions_df <- disease_params_table() flextable(assumptions_df) %>% set_header_labels( Condition = "Condition", Generation_Time = "Generation Time", Incubation_Period = "Incubation Period", Rt_Prior = "Rt Prior", Reporting_Delay = "Reporting Delay", In_Plain_Language = "What This Means", Reference = "Reference" ) %>% theme_vanilla() %>% autofit() %>% fontsize(size = 9, part = "all") %>% fontsize(size = 8, j = "Reference", part = "body") %>% fontsize(size = 8, j = "In_Plain_Language", part = "body") %>% bold(part = "header") %>% width(j = "In_Plain_Language", width = 3.5) %>% width(j = "Reference", width = 2.5) %>% add_footer_lines("Generation time ≈ serial interval where direct measurement is unavailable. Incubation period = time from exposure to symptom onset. All distributions are Gamma. Reporting delay: default LogNormal(mean=3, sd=2, max=15) when empirical delay data is unavailable.") ``` ::: {.callout-tip} ### Reading the parameters | Parameter | What it controls | Why it matters | |---|---|---| | **Generation time** | Average time between successive infections in a chain (≈ serial interval) | Determines how fast $R_t$ changes translate into case-count changes | | **Incubation period** | Time from infection/exposure to symptom onset | Affects the delay between infection dynamics and observed cases | | **$R_t$ prior** | Starting assumption for reproduction number | Weakly informative — the model quickly learns from data | | **Reporting delay** | Time from symptom onset to NMC notification | Accounts for the surveillance lag between illness and official capture | | **What This Means** | Plain-language summary of all parameters for this condition | Helps non-technical readers understand what the model assumes | | **Reference** | Primary literature source | Used for validation — update if newer estimates become available | ::: ::: {.callout-important} ### How to adjust parameters To update parameter values for a condition: 1. Open `R/epinow2_functions.r` 2. Find the condition in `disease_params()` (search for the condition name in lowercase) 3. Change `mean`, `sd`, `max` values to match your preferred literature source 4. Update the `reference` string to cite the new source 5. Re-run: `Rscript R/run_forecasts.r` 6. Re-render: `quarto render by-forecasts/index.qmd` Example — changing measles generation time to use a new study: ```r "fever-rash" = , "measles" = list( generation_time = EpiNow2::Gamma(mean = 11.0, sd = 2.5, max = 20), # ↑ changed from 11.7 to 11.0 based on New Study (2026) incubation_period = EpiNow2::Gamma(mean = 12.5, sd = 2.3, max = 21), rt_prior = list(mean = 2, sd = 1), reference = "New Study (2026) Journal Name DOI:..." ), ``` ::: --- ## Methods & Technical Details {#methods} ::: {.callout-note collapse="true"} ### Technical details (click to expand) **Model:** EpiNow2 v`r as.character(packageVersion("EpiNow2"))` – a Bayesian semi-mechanistic model that jointly estimates the effective reproduction number ($R_t$), infection incidence, and reporting patterns. The model is fitted using Hamiltonian Monte Carlo via CmdStan. A **day-of-week effect** is enabled (`obs_opts(week_effect = TRUE)`) to account for lower weekend reporting. **Display smoothing:** All custom forecast plots apply a **7-day rolling average** to both the observed surveillance series and the EpiNow2 median / credible-interval outputs. This removes residual day-of-week jaggedness and highlights the underlying epidemic trend. **Reporting delay:** When empirical symptom→notification dates are available in the NMC data, the delay distribution is estimated via `EpiNow2::estimate_delay()` (capped at 15 days, max 100 observations). Otherwise a default `LogNormal(mean=3, sd=2, max=15)` prior is used. **Generation time & incubation period:** Condition-specific Gamma distributions from published peer-reviewed literature. Key sources: Lessler et al. (2009) for measles/rubella; Vink et al. (2014) for serial intervals; Azman et al. (2013) for cholera; Park et al. (2023) for COVID-19 Omicron estimates. See the parameter table above for full citations. These are fixed inputs — not estimated from the data. **Priors:** $R_t$ priors are weakly informative: `Normal(2, 1)` for directly-transmitted pathogens, `Normal(1.5, 0.5)` for vector-borne / lower-Rt conditions. The model quickly learns from data. **Forecast horizon:** `r if (!is.null(meta)) meta$horizon else 28` days (4 weeks) ahead from the last observation date. **History window:** `r if (!is.null(meta)) meta$lookback else 360` days of surveillance data fed to each model run. **Completeness / Timeliness:** The NMC is a passive surveillance system; under-reporting varies by condition and province. Forecasts represent **reported** cases, not true incidence. **Rolling forecast animation:** Re-runs the model `r if (!is.null(meta)) meta$n_rewinds else 8` times (weekly snapshots), each using data up to that snapshot date, to visualise how the forecast outlook evolved over time. ::: --- <footer> *Source code:* [`R/epinow2_functions.r`](../R/epinow2_functions.r) · [`R/run_forecasts.r`](../R/run_forecasts.r) · *Data pipeline:* [`R/prepare_dashboard_data.r`](../R/prepare_dashboard_data.r) </footer>